Reproduce “Polynomial Residual Networks are Generically Identifiable” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Isabel Dahlgren, Giovanni Luca Marchetti, Kathlén Kohn
The authors analyze when constant-width residual networks with polynomial activations have unique parameter representations of their functions. They characterize sets of weights and biases yielding the same function, using exact algebraic arguments. For generic parameters, degree-at-least-five activations with generic coefficients give a singleton fiber, even over the complex numbers; this does not cover exceptional parameters or other network classes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.